OpenRouter Introduces Fusion Hybrid Model: Multiple Models Combined, Half the Price of Fable 5

According to Dynamic Insights monitoring, OpenRouter has launched the Fusion composite model interface, which supports parallel distribution of user prompts to multiple large language models. It then utilizes a referee model and a synthesis model to output the final answer. In the DRACO deep research benchmark published by Perplexity AI, various model combinations based on Fusion demonstrated a suppression effect on traditional single models. Across various configurations, pairing Fable 5 with GPT-5.5 in a panel and synthesizing under Opus 4.8 achieved the highest score of 69.0%, significantly surpassing the 65.3% score of the Fable 5 single model run. The tests indicate that due to the underlying training and logical differences of models from different manufacturers, the diversity of multi-model hybridization can provide a stronger complementary perspective in complex tasks.

Even the same model engaging in a self-play strategy showed a significant improvement. By composing Opus 4.8 with itself into a dual panel for self-synthesis, the score increased from an independent run of 58.8% to 65.5%. For cost-effective requirements, a low-cost panel composed of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro also achieved a score of 64.7% in synthesis. With a halved invocation cost, the gap with Fable 5 narrowed to 0.6 percentage points.

The benchmark test consists of 100 complex research tasks across 10 dimensions. To prevent models from inadvertently accessing online evaluation standards during retrieval, OpenRouter has configured a filtering retrieval domain list on the server side. Fusion is now open to API users with the default identifier openrouter/fusion, and users can also customize the model panels participating in the synthesis on the web.